**LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models** **Authors:** Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyuan Luo, Zhangchi Feng, Yongqiang Ma **Institution:** School of Computer Science and Engineering, Beihang University; School of Software and Microelectronics, Peking University **GitHub Link:** <https://github.com/hiyouga/LLaMA-Factory> **Abstract:** Efficient fine-tuning is crucial for adapting large language models (LLMs) to downstream tasks, but implementing these methods on different models can be challenging. LlamaFactory is a unified framework that integrates cutting-edge efficient training methods, providing a solution for customizing the fine-tuning of over 100 LLMs without coding through its web UI, LlamaBOARD. The framework supports various training approaches, including generative pre-training, supervised fine-tuning, reinforcement learning from human feedback, and direct preference optimization. It has been validated on language modeling and text generation tasks, demonstrating efficiency and effectiveness. **Key Features:** - **Unified Framework:** Integrates multiple efficient training methods. - **Web UI (LlamaBOARD):** enables codeless customization and monitoring. - **Scalability:** Supports a wide range of models and datasets. - **Efficient Training Techniques:** Includes freeze-tuning, gradient low-rank projection, BAdam, LoRA, QLoRA, DoRA, PiSSA, and more. - **Data Processing:** Standardizes datasets for efficient processing. - **Model Sharing RLHF:** enables RLHF training on consumer devices. - **Distributed Training:** supports advanced parallelism strategies. **Empirical Study:** - **Training Efficiency:** Compared memory usage, throughput, and perplexity across different fine-tuning methods. - **Effectiveness on Downstream Tasks:** Evaluated performance on text generation tasks using ROUGE scores. **Conclusion:** LlamaFactory is a comprehensive and efficient framework for fine-tuning LLMs, offering a user-friendly interface and robust training techniques. It has been widely adopted and further development is planned to support more modalities and advanced training strategies.**LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models** **Authors:** Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyuan Luo, Zhangchi Feng, Yongqiang Ma **Institution:** School of Computer Science and Engineering, Beihang University; School of Software and Microelectronics, Peking University **GitHub Link:** <https://github.com/hiyouga/LLaMA-Factory> **Abstract:** Efficient fine-tuning is crucial for adapting large language models (LLMs) to downstream tasks, but implementing these methods on different models can be challenging. LlamaFactory is a unified framework that integrates cutting-edge efficient training methods, providing a solution for customizing the fine-tuning of over 100 LLMs without coding through its web UI, LlamaBOARD. The framework supports various training approaches, including generative pre-training, supervised fine-tuning, reinforcement learning from human feedback, and direct preference optimization. It has been validated on language modeling and text generation tasks, demonstrating efficiency and effectiveness. **Key Features:** - **Unified Framework:** Integrates multiple efficient training methods. - **Web UI (LlamaBOARD):** enables codeless customization and monitoring. - **Scalability:** Supports a wide range of models and datasets. - **Efficient Training Techniques:** Includes freeze-tuning, gradient low-rank projection, BAdam, LoRA, QLoRA, DoRA, PiSSA, and more. - **Data Processing:** Standardizes datasets for efficient processing. - **Model Sharing RLHF:** enables RLHF training on consumer devices. - **Distributed Training:** supports advanced parallelism strategies. **Empirical Study:** - **Training Efficiency:** Compared memory usage, throughput, and perplexity across different fine-tuning methods. - **Effectiveness on Downstream Tasks:** Evaluated performance on text generation tasks using ROUGE scores. **Conclusion:** LlamaFactory is a comprehensive and efficient framework for fine-tuning LLMs, offering a user-friendly interface and robust training techniques. It has been widely adopted and further development is planned to support more modalities and advanced training strategies.